Gemini 4 Hydrogen by Google

Overview

Gemini 4 Hydrogen is a multimodal large language model from Google DeepMind. It takes text plus optional images, audio, video, and documents, then returns text responses. It’s designed for everyday coding, long-document work, and agent-style tasks that need many steps. The 1,048,576-token unified context window helps you keep big repos, long PDFs, or long chats in one thread.

What you can build

• Code review and refactors across multiple files
• Bug triage and step-by-step debugging plans
• Test generation, fix suggestions, and PR summaries
• Document Q&A over PDFs, DOCX, PPTX, and spreadsheets
• Extraction of key fields from contracts or reports
• Meeting or lecture transcription, summaries, and translations from audio
• Video understanding tasks like scene Q&A and timeline summaries
• Agent workflows that switch between planning and execution across many turns

Inputs

• A written instruction or question in plain text.
• Optional chat-style turns (a list of messages with roles and content) when you want multi-turn behavior instead of a single prompt.
• Optional system-level instructions that set rules and tone for the whole run.
• Optional image files in PNG, JPEG, WebP, HEIC, or HEIF format. Images and document files share the same per-prompt item pool.
• Optional document files such as PDF, TXT, DOC, DOCX, PPT, PPTX, XLS, XLSX, and OpenDocument formats. A single document can be up to 3,000 pages.
• Optional audio files such as MP3, WAV, FLAC, OGG, M4A, AAC, PCM, or WebM audio. A prompt can include up to about 8.4 hours of audio, or up to 1 million tokens.
• Optional video files such as MP4, MOV (QuickTime), MPEG, WebM, WMV, FLV, or 3GPP. You can attach up to 10 videos per prompt. Video length is roughly up to 45 minutes if audio is present, or about 1 hour without audio.
• Optional grounding mode that uses either Google Search or Google Maps. This listing treats these as mutually exclusive modes.
• Optional reasoning depth selection with 4 levels (low, medium, high, extra high). Higher levels spend more tokens on harder tasks.
• Optional response length cap. In typical foreground runs, a single response segment can produce up to 65,536 tokens.
• Optional identifiers for user and session, when you want the platform to keep a stable chat history.
• Optional continuation token from a prior run, when you want to continue a long response across segments.

Outputs

The model returns a text response. The response can be plain text, or structured JSON when you request a specific schema and the platform supports it. When you enable grounding, the answer is based on retrieved Search or Maps context and may include grounded references in the response text.

Recommended settings

• Use low thinking for extraction, classification, and simple rewrites.
• Use medium thinking for routine chat, summarization, and standard coding questions.
• Use high thinking for repo-level debugging, longer plans, and tricky reasoning.
• Use extra high thinking for hard agent tasks where you expect many intermediate steps.
• Keep the response length cap close to what you need. Large caps can increase token usage.

Limitations

• The model’s stated knowledge cutoff is March 2026. Verify facts that depend on newer events.
• It can hallucinate. Treat outputs as drafts unless you can check sources.
• OCR for scanned PDFs is not used by default. Low-quality scans or photos of pages can reduce accuracy.
• Sampling controls are fixed in some deployments. Temperature, top-p, and top-k may ignore custom values.
• A single foreground response segment is typically limited to 65,536 output tokens. Longer answers may require continuation.
• Fine-tuning is not supported on this model.
• Some advanced agent features can vary by integration. This listing does not rely on user-defined custom function schemas.
• Large attachments still have practical limits. Public URL fetches can be constrained to small file sizes compared to Cloud Storage.

Safety & compliance

Follow Google’s Gemini usage policies and the platform’s acceptable use rules. Don’t request instructions for wrongdoing, weaponization, or harmful activities. Expect safety filtering and refusals on disallowed requests.

Google reports targeted safety assessments for misuse areas like CBRN and cybersecurity, plus ongoing mitigations and monitoring. You should still add your own safeguards for high-stakes uses, including human review and external verification.